2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * GENERAL RULES ON VM_PAGE MANIPULATION
66 * - A page queue lock is required when adding or removing a page from a
67 * page queue regardless of other locks or the busy state of a page.
69 * * In general, no thread besides the page daemon can acquire or
70 * hold more than one page queue lock at a time.
72 * * The page daemon can acquire and hold any pair of page queue
75 * - The object lock is required when inserting or removing
76 * pages from an object (vm_page_insert() or vm_page_remove()).
81 * Resident memory management module.
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
89 #include <sys/param.h>
90 #include <sys/systm.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/linker.h>
95 #include <sys/malloc.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
102 #include <sys/sysctl.h>
103 #include <sys/vmmeter.h>
104 #include <sys/vnode.h>
108 #include <vm/vm_param.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_object.h>
111 #include <vm/vm_page.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_phys.h>
115 #include <vm/vm_radix.h>
116 #include <vm/vm_reserv.h>
117 #include <vm/vm_extern.h>
119 #include <vm/uma_int.h>
121 #include <machine/md_var.h>
124 * Associated with page of user-allocatable memory is a
128 struct vm_domain vm_dom[MAXMEMDOM];
129 struct mtx_padalign vm_page_queue_free_mtx;
131 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
133 vm_page_t vm_page_array;
134 long vm_page_array_size;
136 int vm_page_zero_count;
138 static int boot_pages = UMA_BOOT_PAGES;
139 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
141 "number of pages allocated for bootstrapping the VM system");
143 static int pa_tryrelock_restart;
144 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
145 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
147 static TAILQ_HEAD(, vm_page) blacklist_head;
148 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
149 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
150 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
153 static uma_zone_t fakepg_zone;
155 static struct vnode *vm_page_alloc_init(vm_page_t m);
156 static void vm_page_cache_turn_free(vm_page_t m);
157 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
158 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
159 static void vm_page_init_fakepg(void *dummy);
160 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
161 vm_pindex_t pindex, vm_page_t mpred);
162 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
165 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
168 vm_page_init_fakepg(void *dummy)
171 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
172 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
175 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
176 #if PAGE_SIZE == 32768
178 CTASSERT(sizeof(u_long) >= 8);
183 * Try to acquire a physical address lock while a pmap is locked. If we
184 * fail to trylock we unlock and lock the pmap directly and cache the
185 * locked pa in *locked. The caller should then restart their loop in case
186 * the virtual to physical mapping has changed.
189 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
196 PA_LOCK_ASSERT(lockpa, MA_OWNED);
197 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
204 atomic_add_int(&pa_tryrelock_restart, 1);
213 * Sets the page size, perhaps based upon the memory
214 * size. Must be called before any use of page-size
215 * dependent functions.
218 vm_set_page_size(void)
220 if (vm_cnt.v_page_size == 0)
221 vm_cnt.v_page_size = PAGE_SIZE;
222 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
223 panic("vm_set_page_size: page size not a power of two");
227 * vm_page_blacklist_next:
229 * Find the next entry in the provided string of blacklist
230 * addresses. Entries are separated by space, comma, or newline.
231 * If an invalid integer is encountered then the rest of the
232 * string is skipped. Updates the list pointer to the next
233 * character, or NULL if the string is exhausted or invalid.
236 vm_page_blacklist_next(char **list, char *end)
241 if (list == NULL || *list == NULL)
249 * If there's no end pointer then the buffer is coming from
250 * the kenv and we know it's null-terminated.
253 end = *list + strlen(*list);
255 /* Ensure that strtoq() won't walk off the end */
257 if (*end == '\n' || *end == ' ' || *end == ',')
260 printf("Blacklist not terminated, skipping\n");
266 for (pos = *list; *pos != '\0'; pos = cp) {
267 bad = strtoq(pos, &cp, 0);
268 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
277 if (*cp == '\0' || ++cp >= end)
281 return (trunc_page(bad));
283 printf("Garbage in RAM blacklist, skipping\n");
289 * vm_page_blacklist_check:
291 * Iterate through the provided string of blacklist addresses, pulling
292 * each entry out of the physical allocator free list and putting it
293 * onto a list for reporting via the vm.page_blacklist sysctl.
296 vm_page_blacklist_check(char *list, char *end)
304 while (next != NULL) {
305 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
307 m = vm_phys_paddr_to_vm_page(pa);
310 mtx_lock(&vm_page_queue_free_mtx);
311 ret = vm_phys_unfree_page(m);
312 mtx_unlock(&vm_page_queue_free_mtx);
314 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
316 printf("Skipping page with pa 0x%jx\n",
323 * vm_page_blacklist_load:
325 * Search for a special module named "ram_blacklist". It'll be a
326 * plain text file provided by the user via the loader directive
330 vm_page_blacklist_load(char **list, char **end)
339 mod = preload_search_by_type("ram_blacklist");
341 ptr = preload_fetch_addr(mod);
342 len = preload_fetch_size(mod);
353 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
360 error = sysctl_wire_old_buffer(req, 0);
363 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
364 TAILQ_FOREACH(m, &blacklist_head, listq) {
365 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
366 (uintmax_t)m->phys_addr);
369 error = sbuf_finish(&sbuf);
375 vm_page_domain_init(struct vm_domain *vmd)
377 struct vm_pagequeue *pq;
380 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
381 "vm inactive pagequeue";
382 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
383 &vm_cnt.v_inactive_count;
384 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
385 "vm active pagequeue";
386 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
387 &vm_cnt.v_active_count;
388 vmd->vmd_page_count = 0;
389 vmd->vmd_free_count = 0;
391 vmd->vmd_oom = FALSE;
393 for (i = 0; i < PQ_COUNT; i++) {
394 pq = &vmd->vmd_pagequeues[i];
395 TAILQ_INIT(&pq->pq_pl);
396 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
397 MTX_DEF | MTX_DUPOK);
404 * Initializes the resident memory module.
406 * Allocates memory for the page cells, and
407 * for the object/offset-to-page hash table headers.
408 * Each page cell is initialized and placed on the free list.
411 vm_page_startup(vm_offset_t vaddr)
414 vm_paddr_t page_range;
419 char *list, *listend;
421 vm_paddr_t biggestsize;
422 vm_paddr_t low_water, high_water;
427 vaddr = round_page(vaddr);
429 for (i = 0; phys_avail[i + 1]; i += 2) {
430 phys_avail[i] = round_page(phys_avail[i]);
431 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
434 low_water = phys_avail[0];
435 high_water = phys_avail[1];
437 for (i = 0; i < vm_phys_nsegs; i++) {
438 if (vm_phys_segs[i].start < low_water)
439 low_water = vm_phys_segs[i].start;
440 if (vm_phys_segs[i].end > high_water)
441 high_water = vm_phys_segs[i].end;
443 for (i = 0; phys_avail[i + 1]; i += 2) {
444 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
446 if (size > biggestsize) {
450 if (phys_avail[i] < low_water)
451 low_water = phys_avail[i];
452 if (phys_avail[i + 1] > high_water)
453 high_water = phys_avail[i + 1];
456 end = phys_avail[biggestone+1];
459 * Initialize the page and queue locks.
461 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
462 for (i = 0; i < PA_LOCK_COUNT; i++)
463 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
464 for (i = 0; i < vm_ndomains; i++)
465 vm_page_domain_init(&vm_dom[i]);
468 * Allocate memory for use when boot strapping the kernel memory
471 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
472 * manually fetch the value.
474 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
475 new_end = end - (boot_pages * UMA_SLAB_SIZE);
476 new_end = trunc_page(new_end);
477 mapped = pmap_map(&vaddr, new_end, end,
478 VM_PROT_READ | VM_PROT_WRITE);
479 bzero((void *)mapped, end - new_end);
480 uma_startup((void *)mapped, boot_pages);
482 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
483 defined(__i386__) || defined(__mips__)
485 * Allocate a bitmap to indicate that a random physical page
486 * needs to be included in a minidump.
488 * The amd64 port needs this to indicate which direct map pages
489 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
491 * However, i386 still needs this workspace internally within the
492 * minidump code. In theory, they are not needed on i386, but are
493 * included should the sf_buf code decide to use them.
496 for (i = 0; dump_avail[i + 1] != 0; i += 2)
497 if (dump_avail[i + 1] > last_pa)
498 last_pa = dump_avail[i + 1];
499 page_range = last_pa / PAGE_SIZE;
500 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
501 new_end -= vm_page_dump_size;
502 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
503 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
504 bzero((void *)vm_page_dump, vm_page_dump_size);
508 * Request that the physical pages underlying the message buffer be
509 * included in a crash dump. Since the message buffer is accessed
510 * through the direct map, they are not automatically included.
512 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
513 last_pa = pa + round_page(msgbufsize);
514 while (pa < last_pa) {
520 * Compute the number of pages of memory that will be available for
521 * use (taking into account the overhead of a page structure per
524 first_page = low_water / PAGE_SIZE;
525 #ifdef VM_PHYSSEG_SPARSE
527 for (i = 0; i < vm_phys_nsegs; i++) {
528 page_range += atop(vm_phys_segs[i].end -
529 vm_phys_segs[i].start);
531 for (i = 0; phys_avail[i + 1] != 0; i += 2)
532 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
533 #elif defined(VM_PHYSSEG_DENSE)
534 page_range = high_water / PAGE_SIZE - first_page;
536 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
541 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
546 * Initialize the mem entry structures now, and put them in the free
549 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
550 mapped = pmap_map(&vaddr, new_end, end,
551 VM_PROT_READ | VM_PROT_WRITE);
552 vm_page_array = (vm_page_t) mapped;
553 #if VM_NRESERVLEVEL > 0
555 * Allocate memory for the reservation management system's data
558 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
560 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
562 * pmap_map on arm64, amd64, and mips can come out of the direct-map,
563 * not kvm like i386, so the pages must be tracked for a crashdump to
564 * include this data. This includes the vm_page_array and the early
565 * UMA bootstrap pages.
567 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
570 phys_avail[biggestone + 1] = new_end;
573 * Add physical memory segments corresponding to the available
576 for (i = 0; phys_avail[i + 1] != 0; i += 2)
577 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
580 * Clear all of the page structures
582 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
583 for (i = 0; i < page_range; i++)
584 vm_page_array[i].order = VM_NFREEORDER;
585 vm_page_array_size = page_range;
588 * Initialize the physical memory allocator.
593 * Add every available physical page that is not blacklisted to
596 vm_cnt.v_page_count = 0;
597 vm_cnt.v_free_count = 0;
598 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
600 last_pa = phys_avail[i + 1];
601 while (pa < last_pa) {
602 vm_phys_add_page(pa);
607 TAILQ_INIT(&blacklist_head);
608 vm_page_blacklist_load(&list, &listend);
609 vm_page_blacklist_check(list, listend);
611 list = kern_getenv("vm.blacklist");
612 vm_page_blacklist_check(list, NULL);
615 #if VM_NRESERVLEVEL > 0
617 * Initialize the reservation management system.
625 vm_page_reference(vm_page_t m)
628 vm_page_aflag_set(m, PGA_REFERENCED);
632 * vm_page_busy_downgrade:
634 * Downgrade an exclusive busy page into a single shared busy page.
637 vm_page_busy_downgrade(vm_page_t m)
641 vm_page_assert_xbusied(m);
645 x &= VPB_BIT_WAITERS;
646 if (atomic_cmpset_rel_int(&m->busy_lock,
647 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
655 * Return a positive value if the page is shared busied, 0 otherwise.
658 vm_page_sbusied(vm_page_t m)
663 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
669 * Shared unbusy a page.
672 vm_page_sunbusy(vm_page_t m)
676 vm_page_assert_sbusied(m);
680 if (VPB_SHARERS(x) > 1) {
681 if (atomic_cmpset_int(&m->busy_lock, x,
686 if ((x & VPB_BIT_WAITERS) == 0) {
687 KASSERT(x == VPB_SHARERS_WORD(1),
688 ("vm_page_sunbusy: invalid lock state"));
689 if (atomic_cmpset_int(&m->busy_lock,
690 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
694 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
695 ("vm_page_sunbusy: invalid lock state for waiters"));
698 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
709 * vm_page_busy_sleep:
711 * Sleep and release the page lock, using the page pointer as wchan.
712 * This is used to implement the hard-path of busying mechanism.
714 * The given page must be locked.
717 vm_page_busy_sleep(vm_page_t m, const char *wmesg)
721 vm_page_lock_assert(m, MA_OWNED);
724 if (x == VPB_UNBUSIED) {
728 if ((x & VPB_BIT_WAITERS) == 0 &&
729 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
733 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
739 * Try to shared busy a page.
740 * If the operation succeeds 1 is returned otherwise 0.
741 * The operation never sleeps.
744 vm_page_trysbusy(vm_page_t m)
750 if ((x & VPB_BIT_SHARED) == 0)
752 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
758 * vm_page_xunbusy_hard:
760 * Called after the first try the exclusive unbusy of a page failed.
761 * It is assumed that the waiters bit is on.
764 vm_page_xunbusy_hard(vm_page_t m)
767 vm_page_assert_xbusied(m);
770 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
778 * Wakeup anyone waiting for the page.
779 * The ownership bits do not change.
781 * The given page must be locked.
784 vm_page_flash(vm_page_t m)
788 vm_page_lock_assert(m, MA_OWNED);
792 if ((x & VPB_BIT_WAITERS) == 0)
794 if (atomic_cmpset_int(&m->busy_lock, x,
795 x & (~VPB_BIT_WAITERS)))
802 * Keep page from being freed by the page daemon
803 * much of the same effect as wiring, except much lower
804 * overhead and should be used only for *very* temporary
805 * holding ("wiring").
808 vm_page_hold(vm_page_t mem)
811 vm_page_lock_assert(mem, MA_OWNED);
816 vm_page_unhold(vm_page_t mem)
819 vm_page_lock_assert(mem, MA_OWNED);
820 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
822 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
823 vm_page_free_toq(mem);
827 * vm_page_unhold_pages:
829 * Unhold each of the pages that is referenced by the given array.
832 vm_page_unhold_pages(vm_page_t *ma, int count)
834 struct mtx *mtx, *new_mtx;
837 for (; count != 0; count--) {
839 * Avoid releasing and reacquiring the same page lock.
841 new_mtx = vm_page_lockptr(*ma);
842 if (mtx != new_mtx) {
856 PHYS_TO_VM_PAGE(vm_paddr_t pa)
860 #ifdef VM_PHYSSEG_SPARSE
861 m = vm_phys_paddr_to_vm_page(pa);
863 m = vm_phys_fictitious_to_vm_page(pa);
865 #elif defined(VM_PHYSSEG_DENSE)
869 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
870 m = &vm_page_array[pi - first_page];
873 return (vm_phys_fictitious_to_vm_page(pa));
875 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
882 * Create a fictitious page with the specified physical address and
883 * memory attribute. The memory attribute is the only the machine-
884 * dependent aspect of a fictitious page that must be initialized.
887 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
891 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
892 vm_page_initfake(m, paddr, memattr);
897 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
900 if ((m->flags & PG_FICTITIOUS) != 0) {
902 * The page's memattr might have changed since the
903 * previous initialization. Update the pmap to the
908 m->phys_addr = paddr;
910 /* Fictitious pages don't use "segind". */
911 m->flags = PG_FICTITIOUS;
912 /* Fictitious pages don't use "order" or "pool". */
913 m->oflags = VPO_UNMANAGED;
914 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
918 pmap_page_set_memattr(m, memattr);
924 * Release a fictitious page.
927 vm_page_putfake(vm_page_t m)
930 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
931 KASSERT((m->flags & PG_FICTITIOUS) != 0,
932 ("vm_page_putfake: bad page %p", m));
933 uma_zfree(fakepg_zone, m);
937 * vm_page_updatefake:
939 * Update the given fictitious page to the specified physical address and
943 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
946 KASSERT((m->flags & PG_FICTITIOUS) != 0,
947 ("vm_page_updatefake: bad page %p", m));
948 m->phys_addr = paddr;
949 pmap_page_set_memattr(m, memattr);
958 vm_page_free(vm_page_t m)
961 m->flags &= ~PG_ZERO;
968 * Free a page to the zerod-pages queue
971 vm_page_free_zero(vm_page_t m)
979 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
980 * array which is not the request page.
983 vm_page_readahead_finish(vm_page_t m)
988 * Since the page is not the requested page, whether
989 * it should be activated or deactivated is not
990 * obvious. Empirical results have shown that
991 * deactivating the page is usually the best choice,
992 * unless the page is wanted by another thread.
995 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
998 vm_page_deactivate(m);
1003 * Free the completely invalid page. Such page state
1004 * occurs due to the short read operation which did
1005 * not covered our page at all, or in case when a read
1015 * vm_page_sleep_if_busy:
1017 * Sleep and release the page queues lock if the page is busied.
1018 * Returns TRUE if the thread slept.
1020 * The given page must be unlocked and object containing it must
1024 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1028 vm_page_lock_assert(m, MA_NOTOWNED);
1029 VM_OBJECT_ASSERT_WLOCKED(m->object);
1031 if (vm_page_busied(m)) {
1033 * The page-specific object must be cached because page
1034 * identity can change during the sleep, causing the
1035 * re-lock of a different object.
1036 * It is assumed that a reference to the object is already
1037 * held by the callers.
1041 VM_OBJECT_WUNLOCK(obj);
1042 vm_page_busy_sleep(m, msg);
1043 VM_OBJECT_WLOCK(obj);
1050 * vm_page_dirty_KBI: [ internal use only ]
1052 * Set all bits in the page's dirty field.
1054 * The object containing the specified page must be locked if the
1055 * call is made from the machine-independent layer.
1057 * See vm_page_clear_dirty_mask().
1059 * This function should only be called by vm_page_dirty().
1062 vm_page_dirty_KBI(vm_page_t m)
1065 /* These assertions refer to this operation by its public name. */
1066 KASSERT((m->flags & PG_CACHED) == 0,
1067 ("vm_page_dirty: page in cache!"));
1068 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1069 ("vm_page_dirty: page is invalid!"));
1070 m->dirty = VM_PAGE_BITS_ALL;
1074 * vm_page_insert: [ internal use only ]
1076 * Inserts the given mem entry into the object and object list.
1078 * The object must be locked.
1081 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1085 VM_OBJECT_ASSERT_WLOCKED(object);
1086 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1087 return (vm_page_insert_after(m, object, pindex, mpred));
1091 * vm_page_insert_after:
1093 * Inserts the page "m" into the specified object at offset "pindex".
1095 * The page "mpred" must immediately precede the offset "pindex" within
1096 * the specified object.
1098 * The object must be locked.
1101 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1108 VM_OBJECT_ASSERT_WLOCKED(object);
1109 KASSERT(m->object == NULL,
1110 ("vm_page_insert_after: page already inserted"));
1111 if (mpred != NULL) {
1112 KASSERT(mpred->object == object,
1113 ("vm_page_insert_after: object doesn't contain mpred"));
1114 KASSERT(mpred->pindex < pindex,
1115 ("vm_page_insert_after: mpred doesn't precede pindex"));
1116 msucc = TAILQ_NEXT(mpred, listq);
1118 msucc = TAILQ_FIRST(&object->memq);
1120 KASSERT(msucc->pindex > pindex,
1121 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1124 * Record the object/offset pair in this page
1132 * Now link into the object's ordered list of backed pages.
1134 if (vm_radix_insert(&object->rtree, m)) {
1139 vm_page_insert_radixdone(m, object, mpred);
1144 * vm_page_insert_radixdone:
1146 * Complete page "m" insertion into the specified object after the
1147 * radix trie hooking.
1149 * The page "mpred" must precede the offset "m->pindex" within the
1152 * The object must be locked.
1155 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1158 VM_OBJECT_ASSERT_WLOCKED(object);
1159 KASSERT(object != NULL && m->object == object,
1160 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1161 if (mpred != NULL) {
1162 KASSERT(mpred->object == object,
1163 ("vm_page_insert_after: object doesn't contain mpred"));
1164 KASSERT(mpred->pindex < m->pindex,
1165 ("vm_page_insert_after: mpred doesn't precede pindex"));
1169 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1171 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1174 * Show that the object has one more resident page.
1176 object->resident_page_count++;
1179 * Hold the vnode until the last page is released.
1181 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1182 vhold(object->handle);
1185 * Since we are inserting a new and possibly dirty page,
1186 * update the object's OBJ_MIGHTBEDIRTY flag.
1188 if (pmap_page_is_write_mapped(m))
1189 vm_object_set_writeable_dirty(object);
1195 * Removes the given mem entry from the object/offset-page
1196 * table and the object page list, but do not invalidate/terminate
1197 * the backing store.
1199 * The object must be locked. The page must be locked if it is managed.
1202 vm_page_remove(vm_page_t m)
1207 if ((m->oflags & VPO_UNMANAGED) == 0)
1208 vm_page_lock_assert(m, MA_OWNED);
1209 if ((object = m->object) == NULL)
1211 VM_OBJECT_ASSERT_WLOCKED(object);
1212 if (vm_page_xbusied(m)) {
1214 if ((m->oflags & VPO_UNMANAGED) != 0 &&
1215 !mtx_owned(vm_page_lockptr(m))) {
1220 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1226 * Now remove from the object's list of backed pages.
1228 vm_radix_remove(&object->rtree, m->pindex);
1229 TAILQ_REMOVE(&object->memq, m, listq);
1232 * And show that the object has one fewer resident page.
1234 object->resident_page_count--;
1237 * The vnode may now be recycled.
1239 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1240 vdrop(object->handle);
1248 * Returns the page associated with the object/offset
1249 * pair specified; if none is found, NULL is returned.
1251 * The object must be locked.
1254 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1257 VM_OBJECT_ASSERT_LOCKED(object);
1258 return (vm_radix_lookup(&object->rtree, pindex));
1262 * vm_page_find_least:
1264 * Returns the page associated with the object with least pindex
1265 * greater than or equal to the parameter pindex, or NULL.
1267 * The object must be locked.
1270 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1274 VM_OBJECT_ASSERT_LOCKED(object);
1275 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1276 m = vm_radix_lookup_ge(&object->rtree, pindex);
1281 * Returns the given page's successor (by pindex) within the object if it is
1282 * resident; if none is found, NULL is returned.
1284 * The object must be locked.
1287 vm_page_next(vm_page_t m)
1291 VM_OBJECT_ASSERT_WLOCKED(m->object);
1292 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1293 next->pindex != m->pindex + 1)
1299 * Returns the given page's predecessor (by pindex) within the object if it is
1300 * resident; if none is found, NULL is returned.
1302 * The object must be locked.
1305 vm_page_prev(vm_page_t m)
1309 VM_OBJECT_ASSERT_WLOCKED(m->object);
1310 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1311 prev->pindex != m->pindex - 1)
1317 * Uses the page mnew as a replacement for an existing page at index
1318 * pindex which must be already present in the object.
1320 * The existing page must not be on a paging queue.
1323 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1325 vm_page_t mold, mpred;
1327 VM_OBJECT_ASSERT_WLOCKED(object);
1330 * This function mostly follows vm_page_insert() and
1331 * vm_page_remove() without the radix, object count and vnode
1332 * dance. Double check such functions for more comments.
1334 mpred = vm_radix_lookup(&object->rtree, pindex);
1335 KASSERT(mpred != NULL,
1336 ("vm_page_replace: replacing page not present with pindex"));
1337 mpred = TAILQ_PREV(mpred, respgs, listq);
1339 KASSERT(mpred->pindex < pindex,
1340 ("vm_page_insert_after: mpred doesn't precede pindex"));
1342 mnew->object = object;
1343 mnew->pindex = pindex;
1344 mold = vm_radix_replace(&object->rtree, mnew);
1345 KASSERT(mold->queue == PQ_NONE,
1346 ("vm_page_replace: mold is on a paging queue"));
1348 /* Detach the old page from the resident tailq. */
1349 TAILQ_REMOVE(&object->memq, mold, listq);
1351 mold->object = NULL;
1352 vm_page_xunbusy(mold);
1354 /* Insert the new page in the resident tailq. */
1356 TAILQ_INSERT_AFTER(&object->memq, mpred, mnew, listq);
1358 TAILQ_INSERT_HEAD(&object->memq, mnew, listq);
1359 if (pmap_page_is_write_mapped(mnew))
1360 vm_object_set_writeable_dirty(object);
1367 * Move the given memory entry from its
1368 * current object to the specified target object/offset.
1370 * Note: swap associated with the page must be invalidated by the move. We
1371 * have to do this for several reasons: (1) we aren't freeing the
1372 * page, (2) we are dirtying the page, (3) the VM system is probably
1373 * moving the page from object A to B, and will then later move
1374 * the backing store from A to B and we can't have a conflict.
1376 * Note: we *always* dirty the page. It is necessary both for the
1377 * fact that we moved it, and because we may be invalidating
1378 * swap. If the page is on the cache, we have to deactivate it
1379 * or vm_page_dirty() will panic. Dirty pages are not allowed
1382 * The objects must be locked.
1385 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1390 VM_OBJECT_ASSERT_WLOCKED(new_object);
1392 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1393 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1394 ("vm_page_rename: pindex already renamed"));
1397 * Create a custom version of vm_page_insert() which does not depend
1398 * by m_prev and can cheat on the implementation aspects of the
1402 m->pindex = new_pindex;
1403 if (vm_radix_insert(&new_object->rtree, m)) {
1409 * The operation cannot fail anymore. The removal must happen before
1410 * the listq iterator is tainted.
1416 /* Return back to the new pindex to complete vm_page_insert(). */
1417 m->pindex = new_pindex;
1418 m->object = new_object;
1420 vm_page_insert_radixdone(m, new_object, mpred);
1426 * Convert all of the given object's cached pages that have a
1427 * pindex within the given range into free pages. If the value
1428 * zero is given for "end", then the range's upper bound is
1429 * infinity. If the given object is backed by a vnode and it
1430 * transitions from having one or more cached pages to none, the
1431 * vnode's hold count is reduced.
1434 vm_page_cache_free(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1439 mtx_lock(&vm_page_queue_free_mtx);
1440 if (__predict_false(vm_radix_is_empty(&object->cache))) {
1441 mtx_unlock(&vm_page_queue_free_mtx);
1444 while ((m = vm_radix_lookup_ge(&object->cache, start)) != NULL) {
1445 if (end != 0 && m->pindex >= end)
1447 vm_radix_remove(&object->cache, m->pindex);
1448 vm_page_cache_turn_free(m);
1450 empty = vm_radix_is_empty(&object->cache);
1451 mtx_unlock(&vm_page_queue_free_mtx);
1452 if (object->type == OBJT_VNODE && empty)
1453 vdrop(object->handle);
1457 * Returns the cached page that is associated with the given
1458 * object and offset. If, however, none exists, returns NULL.
1460 * The free page queue must be locked.
1462 static inline vm_page_t
1463 vm_page_cache_lookup(vm_object_t object, vm_pindex_t pindex)
1466 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1467 return (vm_radix_lookup(&object->cache, pindex));
1471 * Remove the given cached page from its containing object's
1472 * collection of cached pages.
1474 * The free page queue must be locked.
1477 vm_page_cache_remove(vm_page_t m)
1480 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1481 KASSERT((m->flags & PG_CACHED) != 0,
1482 ("vm_page_cache_remove: page %p is not cached", m));
1483 vm_radix_remove(&m->object->cache, m->pindex);
1485 vm_cnt.v_cache_count--;
1489 * Transfer all of the cached pages with offset greater than or
1490 * equal to 'offidxstart' from the original object's cache to the
1491 * new object's cache. However, any cached pages with offset
1492 * greater than or equal to the new object's size are kept in the
1493 * original object. Initially, the new object's cache must be
1494 * empty. Offset 'offidxstart' in the original object must
1495 * correspond to offset zero in the new object.
1497 * The new object must be locked.
1500 vm_page_cache_transfer(vm_object_t orig_object, vm_pindex_t offidxstart,
1501 vm_object_t new_object)
1506 * Insertion into an object's collection of cached pages
1507 * requires the object to be locked. In contrast, removal does
1510 VM_OBJECT_ASSERT_WLOCKED(new_object);
1511 KASSERT(vm_radix_is_empty(&new_object->cache),
1512 ("vm_page_cache_transfer: object %p has cached pages",
1514 mtx_lock(&vm_page_queue_free_mtx);
1515 while ((m = vm_radix_lookup_ge(&orig_object->cache,
1516 offidxstart)) != NULL) {
1518 * Transfer all of the pages with offset greater than or
1519 * equal to 'offidxstart' from the original object's
1520 * cache to the new object's cache.
1522 if ((m->pindex - offidxstart) >= new_object->size)
1524 vm_radix_remove(&orig_object->cache, m->pindex);
1525 /* Update the page's object and offset. */
1526 m->object = new_object;
1527 m->pindex -= offidxstart;
1528 if (vm_radix_insert(&new_object->cache, m))
1529 vm_page_cache_turn_free(m);
1531 mtx_unlock(&vm_page_queue_free_mtx);
1535 * Returns TRUE if a cached page is associated with the given object and
1536 * offset, and FALSE otherwise.
1538 * The object must be locked.
1541 vm_page_is_cached(vm_object_t object, vm_pindex_t pindex)
1546 * Insertion into an object's collection of cached pages requires the
1547 * object to be locked. Therefore, if the object is locked and the
1548 * object's collection is empty, there is no need to acquire the free
1549 * page queues lock in order to prove that the specified page doesn't
1552 VM_OBJECT_ASSERT_WLOCKED(object);
1553 if (__predict_true(vm_object_cache_is_empty(object)))
1555 mtx_lock(&vm_page_queue_free_mtx);
1556 m = vm_page_cache_lookup(object, pindex);
1557 mtx_unlock(&vm_page_queue_free_mtx);
1564 * Allocate and return a page that is associated with the specified
1565 * object and offset pair. By default, this page is exclusive busied.
1567 * The caller must always specify an allocation class.
1569 * allocation classes:
1570 * VM_ALLOC_NORMAL normal process request
1571 * VM_ALLOC_SYSTEM system *really* needs a page
1572 * VM_ALLOC_INTERRUPT interrupt time request
1574 * optional allocation flags:
1575 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1576 * intends to allocate
1577 * VM_ALLOC_IFCACHED return page only if it is cached
1578 * VM_ALLOC_IFNOTCACHED return NULL, do not reactivate if the page
1580 * VM_ALLOC_NOBUSY do not exclusive busy the page
1581 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1582 * VM_ALLOC_NOOBJ page is not associated with an object and
1583 * should not be exclusive busy
1584 * VM_ALLOC_SBUSY shared busy the allocated page
1585 * VM_ALLOC_WIRED wire the allocated page
1586 * VM_ALLOC_ZERO prefer a zeroed page
1588 * This routine may not sleep.
1591 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1593 struct vnode *vp = NULL;
1594 vm_object_t m_object;
1596 int flags, req_class;
1598 mpred = 0; /* XXX: pacify gcc */
1599 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1600 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1601 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1602 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1603 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1606 VM_OBJECT_ASSERT_WLOCKED(object);
1608 req_class = req & VM_ALLOC_CLASS_MASK;
1611 * The page daemon is allowed to dig deeper into the free page list.
1613 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1614 req_class = VM_ALLOC_SYSTEM;
1616 if (object != NULL) {
1617 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1618 KASSERT(mpred == NULL || mpred->pindex != pindex,
1619 ("vm_page_alloc: pindex already allocated"));
1623 * The page allocation request can came from consumers which already
1624 * hold the free page queue mutex, like vm_page_insert() in
1627 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1628 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
1629 (req_class == VM_ALLOC_SYSTEM &&
1630 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
1631 (req_class == VM_ALLOC_INTERRUPT &&
1632 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0)) {
1634 * Allocate from the free queue if the number of free pages
1635 * exceeds the minimum for the request class.
1637 if (object != NULL &&
1638 (m = vm_page_cache_lookup(object, pindex)) != NULL) {
1639 if ((req & VM_ALLOC_IFNOTCACHED) != 0) {
1640 mtx_unlock(&vm_page_queue_free_mtx);
1643 if (vm_phys_unfree_page(m))
1644 vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0);
1645 #if VM_NRESERVLEVEL > 0
1646 else if (!vm_reserv_reactivate_page(m))
1650 panic("vm_page_alloc: cache page %p is missing"
1651 " from the free queue", m);
1652 } else if ((req & VM_ALLOC_IFCACHED) != 0) {
1653 mtx_unlock(&vm_page_queue_free_mtx);
1655 #if VM_NRESERVLEVEL > 0
1656 } else if (object == NULL || (object->flags & (OBJ_COLORED |
1657 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1658 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1662 m = vm_phys_alloc_pages(object != NULL ?
1663 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1664 #if VM_NRESERVLEVEL > 0
1665 if (m == NULL && vm_reserv_reclaim_inactive()) {
1666 m = vm_phys_alloc_pages(object != NULL ?
1667 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1674 * Not allocatable, give up.
1676 mtx_unlock(&vm_page_queue_free_mtx);
1677 atomic_add_int(&vm_pageout_deficit,
1678 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1679 pagedaemon_wakeup();
1684 * At this point we had better have found a good page.
1686 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1687 KASSERT(m->queue == PQ_NONE,
1688 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1689 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1690 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1691 KASSERT(!vm_page_sbusied(m),
1692 ("vm_page_alloc: page %p is busy", m));
1693 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1694 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1695 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1696 pmap_page_get_memattr(m)));
1697 if ((m->flags & PG_CACHED) != 0) {
1698 KASSERT((m->flags & PG_ZERO) == 0,
1699 ("vm_page_alloc: cached page %p is PG_ZERO", m));
1700 KASSERT(m->valid != 0,
1701 ("vm_page_alloc: cached page %p is invalid", m));
1702 if (m->object == object && m->pindex == pindex)
1703 vm_cnt.v_reactivated++;
1706 m_object = m->object;
1707 vm_page_cache_remove(m);
1708 if (m_object->type == OBJT_VNODE &&
1709 vm_object_cache_is_empty(m_object))
1710 vp = m_object->handle;
1712 KASSERT(m->valid == 0,
1713 ("vm_page_alloc: free page %p is valid", m));
1714 vm_phys_freecnt_adj(m, -1);
1715 if ((m->flags & PG_ZERO) != 0)
1716 vm_page_zero_count--;
1718 mtx_unlock(&vm_page_queue_free_mtx);
1721 * Initialize the page. Only the PG_ZERO flag is inherited.
1724 if ((req & VM_ALLOC_ZERO) != 0)
1727 if ((req & VM_ALLOC_NODUMP) != 0)
1731 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1733 m->busy_lock = VPB_UNBUSIED;
1734 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1735 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1736 if ((req & VM_ALLOC_SBUSY) != 0)
1737 m->busy_lock = VPB_SHARERS_WORD(1);
1738 if (req & VM_ALLOC_WIRED) {
1740 * The page lock is not required for wiring a page until that
1741 * page is inserted into the object.
1743 atomic_add_int(&vm_cnt.v_wire_count, 1);
1748 if (object != NULL) {
1749 if (vm_page_insert_after(m, object, pindex, mpred)) {
1750 /* See the comment below about hold count. */
1753 pagedaemon_wakeup();
1754 if (req & VM_ALLOC_WIRED) {
1755 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1759 m->oflags = VPO_UNMANAGED;
1764 /* Ignore device objects; the pager sets "memattr" for them. */
1765 if (object->memattr != VM_MEMATTR_DEFAULT &&
1766 (object->flags & OBJ_FICTITIOUS) == 0)
1767 pmap_page_set_memattr(m, object->memattr);
1772 * The following call to vdrop() must come after the above call
1773 * to vm_page_insert() in case both affect the same object and
1774 * vnode. Otherwise, the affected vnode's hold count could
1775 * temporarily become zero.
1781 * Don't wakeup too often - wakeup the pageout daemon when
1782 * we would be nearly out of memory.
1784 if (vm_paging_needed())
1785 pagedaemon_wakeup();
1791 vm_page_alloc_contig_vdrop(struct spglist *lst)
1794 while (!SLIST_EMPTY(lst)) {
1795 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1796 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1801 * vm_page_alloc_contig:
1803 * Allocate a contiguous set of physical pages of the given size "npages"
1804 * from the free lists. All of the physical pages must be at or above
1805 * the given physical address "low" and below the given physical address
1806 * "high". The given value "alignment" determines the alignment of the
1807 * first physical page in the set. If the given value "boundary" is
1808 * non-zero, then the set of physical pages cannot cross any physical
1809 * address boundary that is a multiple of that value. Both "alignment"
1810 * and "boundary" must be a power of two.
1812 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1813 * then the memory attribute setting for the physical pages is configured
1814 * to the object's memory attribute setting. Otherwise, the memory
1815 * attribute setting for the physical pages is configured to "memattr",
1816 * overriding the object's memory attribute setting. However, if the
1817 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1818 * memory attribute setting for the physical pages cannot be configured
1819 * to VM_MEMATTR_DEFAULT.
1821 * The caller must always specify an allocation class.
1823 * allocation classes:
1824 * VM_ALLOC_NORMAL normal process request
1825 * VM_ALLOC_SYSTEM system *really* needs a page
1826 * VM_ALLOC_INTERRUPT interrupt time request
1828 * optional allocation flags:
1829 * VM_ALLOC_NOBUSY do not exclusive busy the page
1830 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1831 * VM_ALLOC_NOOBJ page is not associated with an object and
1832 * should not be exclusive busy
1833 * VM_ALLOC_SBUSY shared busy the allocated page
1834 * VM_ALLOC_WIRED wire the allocated page
1835 * VM_ALLOC_ZERO prefer a zeroed page
1837 * This routine may not sleep.
1840 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1841 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1842 vm_paddr_t boundary, vm_memattr_t memattr)
1845 struct spglist deferred_vdrop_list;
1846 vm_page_t m, m_tmp, m_ret;
1850 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1851 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1852 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1853 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1854 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1856 if (object != NULL) {
1857 VM_OBJECT_ASSERT_WLOCKED(object);
1858 KASSERT(object->type == OBJT_PHYS,
1859 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1862 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1863 req_class = req & VM_ALLOC_CLASS_MASK;
1866 * The page daemon is allowed to dig deeper into the free page list.
1868 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1869 req_class = VM_ALLOC_SYSTEM;
1871 SLIST_INIT(&deferred_vdrop_list);
1872 mtx_lock(&vm_page_queue_free_mtx);
1873 if (vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1874 vm_cnt.v_free_reserved || (req_class == VM_ALLOC_SYSTEM &&
1875 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages +
1876 vm_cnt.v_interrupt_free_min) || (req_class == VM_ALLOC_INTERRUPT &&
1877 vm_cnt.v_free_count + vm_cnt.v_cache_count >= npages)) {
1878 #if VM_NRESERVLEVEL > 0
1880 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1881 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1882 low, high, alignment, boundary)) == NULL)
1884 m_ret = vm_phys_alloc_contig(npages, low, high,
1885 alignment, boundary);
1887 mtx_unlock(&vm_page_queue_free_mtx);
1888 atomic_add_int(&vm_pageout_deficit, npages);
1889 pagedaemon_wakeup();
1893 for (m = m_ret; m < &m_ret[npages]; m++) {
1894 drop = vm_page_alloc_init(m);
1897 * Enqueue the vnode for deferred vdrop().
1899 m->plinks.s.pv = drop;
1900 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1905 #if VM_NRESERVLEVEL > 0
1906 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1911 mtx_unlock(&vm_page_queue_free_mtx);
1916 * Initialize the pages. Only the PG_ZERO flag is inherited.
1919 if ((req & VM_ALLOC_ZERO) != 0)
1921 if ((req & VM_ALLOC_NODUMP) != 0)
1923 if ((req & VM_ALLOC_WIRED) != 0)
1924 atomic_add_int(&vm_cnt.v_wire_count, npages);
1925 if (object != NULL) {
1926 if (object->memattr != VM_MEMATTR_DEFAULT &&
1927 memattr == VM_MEMATTR_DEFAULT)
1928 memattr = object->memattr;
1930 for (m = m_ret; m < &m_ret[npages]; m++) {
1932 m->flags = (m->flags | PG_NODUMP) & flags;
1933 m->busy_lock = VPB_UNBUSIED;
1934 if (object != NULL) {
1935 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1936 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1937 if ((req & VM_ALLOC_SBUSY) != 0)
1938 m->busy_lock = VPB_SHARERS_WORD(1);
1940 if ((req & VM_ALLOC_WIRED) != 0)
1942 /* Unmanaged pages don't use "act_count". */
1943 m->oflags = VPO_UNMANAGED;
1944 if (object != NULL) {
1945 if (vm_page_insert(m, object, pindex)) {
1946 vm_page_alloc_contig_vdrop(
1947 &deferred_vdrop_list);
1948 if (vm_paging_needed())
1949 pagedaemon_wakeup();
1950 if ((req & VM_ALLOC_WIRED) != 0)
1951 atomic_subtract_int(&vm_cnt.v_wire_count,
1953 for (m_tmp = m, m = m_ret;
1954 m < &m_ret[npages]; m++) {
1955 if ((req & VM_ALLOC_WIRED) != 0)
1965 if (memattr != VM_MEMATTR_DEFAULT)
1966 pmap_page_set_memattr(m, memattr);
1969 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1970 if (vm_paging_needed())
1971 pagedaemon_wakeup();
1976 * Initialize a page that has been freshly dequeued from a freelist.
1977 * The caller has to drop the vnode returned, if it is not NULL.
1979 * This function may only be used to initialize unmanaged pages.
1981 * To be called with vm_page_queue_free_mtx held.
1983 static struct vnode *
1984 vm_page_alloc_init(vm_page_t m)
1987 vm_object_t m_object;
1989 KASSERT(m->queue == PQ_NONE,
1990 ("vm_page_alloc_init: page %p has unexpected queue %d",
1992 KASSERT(m->wire_count == 0,
1993 ("vm_page_alloc_init: page %p is wired", m));
1994 KASSERT(m->hold_count == 0,
1995 ("vm_page_alloc_init: page %p is held", m));
1996 KASSERT(!vm_page_sbusied(m),
1997 ("vm_page_alloc_init: page %p is busy", m));
1998 KASSERT(m->dirty == 0,
1999 ("vm_page_alloc_init: page %p is dirty", m));
2000 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2001 ("vm_page_alloc_init: page %p has unexpected memattr %d",
2002 m, pmap_page_get_memattr(m)));
2003 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2005 if ((m->flags & PG_CACHED) != 0) {
2006 KASSERT((m->flags & PG_ZERO) == 0,
2007 ("vm_page_alloc_init: cached page %p is PG_ZERO", m));
2009 m_object = m->object;
2010 vm_page_cache_remove(m);
2011 if (m_object->type == OBJT_VNODE &&
2012 vm_object_cache_is_empty(m_object))
2013 drop = m_object->handle;
2015 KASSERT(m->valid == 0,
2016 ("vm_page_alloc_init: free page %p is valid", m));
2017 vm_phys_freecnt_adj(m, -1);
2018 if ((m->flags & PG_ZERO) != 0)
2019 vm_page_zero_count--;
2025 * vm_page_alloc_freelist:
2027 * Allocate a physical page from the specified free page list.
2029 * The caller must always specify an allocation class.
2031 * allocation classes:
2032 * VM_ALLOC_NORMAL normal process request
2033 * VM_ALLOC_SYSTEM system *really* needs a page
2034 * VM_ALLOC_INTERRUPT interrupt time request
2036 * optional allocation flags:
2037 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
2038 * intends to allocate
2039 * VM_ALLOC_WIRED wire the allocated page
2040 * VM_ALLOC_ZERO prefer a zeroed page
2042 * This routine may not sleep.
2045 vm_page_alloc_freelist(int flind, int req)
2052 req_class = req & VM_ALLOC_CLASS_MASK;
2055 * The page daemon is allowed to dig deeper into the free page list.
2057 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2058 req_class = VM_ALLOC_SYSTEM;
2061 * Do not allocate reserved pages unless the req has asked for it.
2063 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
2064 if (vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_free_reserved ||
2065 (req_class == VM_ALLOC_SYSTEM &&
2066 vm_cnt.v_free_count + vm_cnt.v_cache_count > vm_cnt.v_interrupt_free_min) ||
2067 (req_class == VM_ALLOC_INTERRUPT &&
2068 vm_cnt.v_free_count + vm_cnt.v_cache_count > 0))
2069 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
2071 mtx_unlock(&vm_page_queue_free_mtx);
2072 atomic_add_int(&vm_pageout_deficit,
2073 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2074 pagedaemon_wakeup();
2078 mtx_unlock(&vm_page_queue_free_mtx);
2081 drop = vm_page_alloc_init(m);
2082 mtx_unlock(&vm_page_queue_free_mtx);
2085 * Initialize the page. Only the PG_ZERO flag is inherited.
2089 if ((req & VM_ALLOC_ZERO) != 0)
2092 if ((req & VM_ALLOC_WIRED) != 0) {
2094 * The page lock is not required for wiring a page that does
2095 * not belong to an object.
2097 atomic_add_int(&vm_cnt.v_wire_count, 1);
2100 /* Unmanaged pages don't use "act_count". */
2101 m->oflags = VPO_UNMANAGED;
2104 if (vm_paging_needed())
2105 pagedaemon_wakeup();
2110 * vm_wait: (also see VM_WAIT macro)
2112 * Sleep until free pages are available for allocation.
2113 * - Called in various places before memory allocations.
2119 mtx_lock(&vm_page_queue_free_mtx);
2120 if (curproc == pageproc) {
2121 vm_pageout_pages_needed = 1;
2122 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2123 PDROP | PSWP, "VMWait", 0);
2125 if (!vm_pages_needed) {
2126 vm_pages_needed = 1;
2127 wakeup(&vm_pages_needed);
2129 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2135 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2137 * Sleep until free pages are available for allocation.
2138 * - Called only in vm_fault so that processes page faulting
2139 * can be easily tracked.
2140 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2141 * processes will be able to grab memory first. Do not change
2142 * this balance without careful testing first.
2148 mtx_lock(&vm_page_queue_free_mtx);
2149 if (!vm_pages_needed) {
2150 vm_pages_needed = 1;
2151 wakeup(&vm_pages_needed);
2153 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2157 struct vm_pagequeue *
2158 vm_page_pagequeue(vm_page_t m)
2161 return (&vm_phys_domain(m)->vmd_pagequeues[m->queue]);
2167 * Remove the given page from its current page queue.
2169 * The page must be locked.
2172 vm_page_dequeue(vm_page_t m)
2174 struct vm_pagequeue *pq;
2176 vm_page_assert_locked(m);
2177 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2179 pq = vm_page_pagequeue(m);
2180 vm_pagequeue_lock(pq);
2182 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2183 vm_pagequeue_cnt_dec(pq);
2184 vm_pagequeue_unlock(pq);
2188 * vm_page_dequeue_locked:
2190 * Remove the given page from its current page queue.
2192 * The page and page queue must be locked.
2195 vm_page_dequeue_locked(vm_page_t m)
2197 struct vm_pagequeue *pq;
2199 vm_page_lock_assert(m, MA_OWNED);
2200 pq = vm_page_pagequeue(m);
2201 vm_pagequeue_assert_locked(pq);
2203 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2204 vm_pagequeue_cnt_dec(pq);
2210 * Add the given page to the specified page queue.
2212 * The page must be locked.
2215 vm_page_enqueue(uint8_t queue, vm_page_t m)
2217 struct vm_pagequeue *pq;
2219 vm_page_lock_assert(m, MA_OWNED);
2220 KASSERT(queue < PQ_COUNT,
2221 ("vm_page_enqueue: invalid queue %u request for page %p",
2223 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2224 vm_pagequeue_lock(pq);
2226 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2227 vm_pagequeue_cnt_inc(pq);
2228 vm_pagequeue_unlock(pq);
2234 * Move the given page to the tail of its current page queue.
2236 * The page must be locked.
2239 vm_page_requeue(vm_page_t m)
2241 struct vm_pagequeue *pq;
2243 vm_page_lock_assert(m, MA_OWNED);
2244 KASSERT(m->queue != PQ_NONE,
2245 ("vm_page_requeue: page %p is not queued", m));
2246 pq = vm_page_pagequeue(m);
2247 vm_pagequeue_lock(pq);
2248 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2249 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2250 vm_pagequeue_unlock(pq);
2254 * vm_page_requeue_locked:
2256 * Move the given page to the tail of its current page queue.
2258 * The page queue must be locked.
2261 vm_page_requeue_locked(vm_page_t m)
2263 struct vm_pagequeue *pq;
2265 KASSERT(m->queue != PQ_NONE,
2266 ("vm_page_requeue_locked: page %p is not queued", m));
2267 pq = vm_page_pagequeue(m);
2268 vm_pagequeue_assert_locked(pq);
2269 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2270 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2276 * Put the specified page on the active list (if appropriate).
2277 * Ensure that act_count is at least ACT_INIT but do not otherwise
2280 * The page must be locked.
2283 vm_page_activate(vm_page_t m)
2287 vm_page_lock_assert(m, MA_OWNED);
2288 if ((queue = m->queue) != PQ_ACTIVE) {
2289 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2290 if (m->act_count < ACT_INIT)
2291 m->act_count = ACT_INIT;
2292 if (queue != PQ_NONE)
2294 vm_page_enqueue(PQ_ACTIVE, m);
2296 KASSERT(queue == PQ_NONE,
2297 ("vm_page_activate: wired page %p is queued", m));
2299 if (m->act_count < ACT_INIT)
2300 m->act_count = ACT_INIT;
2305 * vm_page_free_wakeup:
2307 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2308 * routine is called when a page has been added to the cache or free
2311 * The page queues must be locked.
2314 vm_page_free_wakeup(void)
2317 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2319 * if pageout daemon needs pages, then tell it that there are
2322 if (vm_pageout_pages_needed &&
2323 vm_cnt.v_cache_count + vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2324 wakeup(&vm_pageout_pages_needed);
2325 vm_pageout_pages_needed = 0;
2328 * wakeup processes that are waiting on memory if we hit a
2329 * high water mark. And wakeup scheduler process if we have
2330 * lots of memory. this process will swapin processes.
2332 if (vm_pages_needed && !vm_page_count_min()) {
2333 vm_pages_needed = 0;
2334 wakeup(&vm_cnt.v_free_count);
2339 * Turn a cached page into a free page, by changing its attributes.
2340 * Keep the statistics up-to-date.
2342 * The free page queue must be locked.
2345 vm_page_cache_turn_free(vm_page_t m)
2348 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2352 KASSERT((m->flags & PG_CACHED) != 0,
2353 ("vm_page_cache_turn_free: page %p is not cached", m));
2354 m->flags &= ~PG_CACHED;
2355 vm_cnt.v_cache_count--;
2356 vm_phys_freecnt_adj(m, 1);
2362 * Returns the given page to the free list,
2363 * disassociating it with any VM object.
2365 * The object must be locked. The page must be locked if it is managed.
2368 vm_page_free_toq(vm_page_t m)
2371 if ((m->oflags & VPO_UNMANAGED) == 0) {
2372 vm_page_lock_assert(m, MA_OWNED);
2373 KASSERT(!pmap_page_is_mapped(m),
2374 ("vm_page_free_toq: freeing mapped page %p", m));
2376 KASSERT(m->queue == PQ_NONE,
2377 ("vm_page_free_toq: unmanaged page %p is queued", m));
2378 PCPU_INC(cnt.v_tfree);
2380 if (vm_page_sbusied(m))
2381 panic("vm_page_free: freeing busy page %p", m);
2384 * Unqueue, then remove page. Note that we cannot destroy
2385 * the page here because we do not want to call the pager's
2386 * callback routine until after we've put the page on the
2387 * appropriate free queue.
2393 * If fictitious remove object association and
2394 * return, otherwise delay object association removal.
2396 if ((m->flags & PG_FICTITIOUS) != 0) {
2403 if (m->wire_count != 0)
2404 panic("vm_page_free: freeing wired page %p", m);
2405 if (m->hold_count != 0) {
2406 m->flags &= ~PG_ZERO;
2407 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2408 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
2409 m->flags |= PG_UNHOLDFREE;
2412 * Restore the default memory attribute to the page.
2414 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2415 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2418 * Insert the page into the physical memory allocator's
2419 * cache/free page queues.
2421 mtx_lock(&vm_page_queue_free_mtx);
2422 vm_phys_freecnt_adj(m, 1);
2423 #if VM_NRESERVLEVEL > 0
2424 if (!vm_reserv_free_page(m))
2428 vm_phys_free_pages(m, 0);
2429 if ((m->flags & PG_ZERO) != 0)
2430 ++vm_page_zero_count;
2432 vm_page_zero_idle_wakeup();
2433 vm_page_free_wakeup();
2434 mtx_unlock(&vm_page_queue_free_mtx);
2441 * Mark this page as wired down by yet
2442 * another map, removing it from paging queues
2445 * If the page is fictitious, then its wire count must remain one.
2447 * The page must be locked.
2450 vm_page_wire(vm_page_t m)
2454 * Only bump the wire statistics if the page is not already wired,
2455 * and only unqueue the page if it is on some queue (if it is unmanaged
2456 * it is already off the queues).
2458 vm_page_lock_assert(m, MA_OWNED);
2459 if ((m->flags & PG_FICTITIOUS) != 0) {
2460 KASSERT(m->wire_count == 1,
2461 ("vm_page_wire: fictitious page %p's wire count isn't one",
2465 if (m->wire_count == 0) {
2466 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
2467 m->queue == PQ_NONE,
2468 ("vm_page_wire: unmanaged page %p is queued", m));
2470 atomic_add_int(&vm_cnt.v_wire_count, 1);
2473 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
2479 * Release one wiring of the specified page, potentially enabling it to be
2480 * paged again. If paging is enabled, then the value of the parameter
2481 * "queue" determines the queue to which the page is added.
2483 * However, unless the page belongs to an object, it is not enqueued because
2484 * it cannot be paged out.
2486 * If a page is fictitious, then its wire count must always be one.
2488 * A managed page must be locked.
2491 vm_page_unwire(vm_page_t m, uint8_t queue)
2494 KASSERT(queue < PQ_COUNT,
2495 ("vm_page_unwire: invalid queue %u request for page %p",
2497 if ((m->oflags & VPO_UNMANAGED) == 0)
2498 vm_page_lock_assert(m, MA_OWNED);
2499 if ((m->flags & PG_FICTITIOUS) != 0) {
2500 KASSERT(m->wire_count == 1,
2501 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
2504 if (m->wire_count > 0) {
2506 if (m->wire_count == 0) {
2507 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
2508 if ((m->oflags & VPO_UNMANAGED) != 0 ||
2511 if (queue == PQ_INACTIVE)
2512 m->flags &= ~PG_WINATCFLS;
2513 vm_page_enqueue(queue, m);
2516 panic("vm_page_unwire: page %p's wire count is zero", m);
2520 * Move the specified page to the inactive queue.
2522 * Many pages placed on the inactive queue should actually go
2523 * into the cache, but it is difficult to figure out which. What
2524 * we do instead, if the inactive target is well met, is to put
2525 * clean pages at the head of the inactive queue instead of the tail.
2526 * This will cause them to be moved to the cache more quickly and
2527 * if not actively re-referenced, reclaimed more quickly. If we just
2528 * stick these pages at the end of the inactive queue, heavy filesystem
2529 * meta-data accesses can cause an unnecessary paging load on memory bound
2530 * processes. This optimization causes one-time-use metadata to be
2531 * reused more quickly.
2533 * Normally athead is 0 resulting in LRU operation. athead is set
2534 * to 1 if we want this page to be 'as if it were placed in the cache',
2535 * except without unmapping it from the process address space.
2537 * The page must be locked.
2540 _vm_page_deactivate(vm_page_t m, int athead)
2542 struct vm_pagequeue *pq;
2545 vm_page_assert_locked(m);
2548 * Ignore if the page is already inactive, unless it is unlikely to be
2551 if ((queue = m->queue) == PQ_INACTIVE && !athead)
2553 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2554 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
2555 /* Avoid multiple acquisitions of the inactive queue lock. */
2556 if (queue == PQ_INACTIVE) {
2557 vm_pagequeue_lock(pq);
2558 vm_page_dequeue_locked(m);
2560 if (queue != PQ_NONE)
2562 m->flags &= ~PG_WINATCFLS;
2563 vm_pagequeue_lock(pq);
2565 m->queue = PQ_INACTIVE;
2567 TAILQ_INSERT_HEAD(&pq->pq_pl, m, plinks.q);
2569 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2570 vm_pagequeue_cnt_inc(pq);
2571 vm_pagequeue_unlock(pq);
2576 * Move the specified page to the inactive queue.
2578 * The page must be locked.
2581 vm_page_deactivate(vm_page_t m)
2584 _vm_page_deactivate(m, 0);
2588 * vm_page_try_to_cache:
2590 * Returns 0 on failure, 1 on success
2593 vm_page_try_to_cache(vm_page_t m)
2596 vm_page_lock_assert(m, MA_OWNED);
2597 VM_OBJECT_ASSERT_WLOCKED(m->object);
2598 if (m->dirty || m->hold_count || m->wire_count ||
2599 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2609 * vm_page_try_to_free()
2611 * Attempt to free the page. If we cannot free it, we do nothing.
2612 * 1 is returned on success, 0 on failure.
2615 vm_page_try_to_free(vm_page_t m)
2618 vm_page_lock_assert(m, MA_OWNED);
2619 if (m->object != NULL)
2620 VM_OBJECT_ASSERT_WLOCKED(m->object);
2621 if (m->dirty || m->hold_count || m->wire_count ||
2622 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
2634 * Put the specified page onto the page cache queue (if appropriate).
2636 * The object and page must be locked.
2639 vm_page_cache(vm_page_t m)
2642 boolean_t cache_was_empty;
2644 vm_page_lock_assert(m, MA_OWNED);
2646 VM_OBJECT_ASSERT_WLOCKED(object);
2647 if (vm_page_busied(m) || (m->oflags & VPO_UNMANAGED) ||
2648 m->hold_count || m->wire_count)
2649 panic("vm_page_cache: attempting to cache busy page");
2650 KASSERT(!pmap_page_is_mapped(m),
2651 ("vm_page_cache: page %p is mapped", m));
2652 KASSERT(m->dirty == 0, ("vm_page_cache: page %p is dirty", m));
2653 if (m->valid == 0 || object->type == OBJT_DEFAULT ||
2654 (object->type == OBJT_SWAP &&
2655 !vm_pager_has_page(object, m->pindex, NULL, NULL))) {
2657 * Hypothesis: A cache-eligible page belonging to a
2658 * default object or swap object but without a backing
2659 * store must be zero filled.
2664 KASSERT((m->flags & PG_CACHED) == 0,
2665 ("vm_page_cache: page %p is already cached", m));
2668 * Remove the page from the paging queues.
2673 * Remove the page from the object's collection of resident
2676 vm_radix_remove(&object->rtree, m->pindex);
2677 TAILQ_REMOVE(&object->memq, m, listq);
2678 object->resident_page_count--;
2681 * Restore the default memory attribute to the page.
2683 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
2684 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
2687 * Insert the page into the object's collection of cached pages
2688 * and the physical memory allocator's cache/free page queues.
2690 m->flags &= ~PG_ZERO;
2691 mtx_lock(&vm_page_queue_free_mtx);
2692 cache_was_empty = vm_radix_is_empty(&object->cache);
2693 if (vm_radix_insert(&object->cache, m)) {
2694 mtx_unlock(&vm_page_queue_free_mtx);
2695 if (object->resident_page_count == 0)
2696 vdrop(object->handle);
2703 * The above call to vm_radix_insert() could reclaim the one pre-
2704 * existing cached page from this object, resulting in a call to
2707 if (!cache_was_empty)
2708 cache_was_empty = vm_radix_is_singleton(&object->cache);
2710 m->flags |= PG_CACHED;
2711 vm_cnt.v_cache_count++;
2712 PCPU_INC(cnt.v_tcached);
2713 #if VM_NRESERVLEVEL > 0
2714 if (!vm_reserv_free_page(m)) {
2718 vm_phys_free_pages(m, 0);
2720 vm_page_free_wakeup();
2721 mtx_unlock(&vm_page_queue_free_mtx);
2724 * Increment the vnode's hold count if this is the object's only
2725 * cached page. Decrement the vnode's hold count if this was
2726 * the object's only resident page.
2728 if (object->type == OBJT_VNODE) {
2729 if (cache_was_empty && object->resident_page_count != 0)
2730 vhold(object->handle);
2731 else if (!cache_was_empty && object->resident_page_count == 0)
2732 vdrop(object->handle);
2739 * Deactivate or do nothing, as appropriate. This routine is used
2740 * by madvise() and vop_stdadvise().
2742 * The object and page must be locked.
2745 vm_page_advise(vm_page_t m, int advice)
2748 vm_page_assert_locked(m);
2749 VM_OBJECT_ASSERT_WLOCKED(m->object);
2750 if (advice == MADV_FREE)
2752 * Mark the page clean. This will allow the page to be freed
2753 * up by the system. However, such pages are often reused
2754 * quickly by malloc() so we do not do anything that would
2755 * cause a page fault if we can help it.
2757 * Specifically, we do not try to actually free the page now
2758 * nor do we try to put it in the cache (which would cause a
2759 * page fault on reuse).
2761 * But we do make the page as freeable as we can without
2762 * actually taking the step of unmapping it.
2765 else if (advice != MADV_DONTNEED)
2769 * Clear any references to the page. Otherwise, the page daemon will
2770 * immediately reactivate the page.
2772 vm_page_aflag_clear(m, PGA_REFERENCED);
2774 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
2778 * Place clean pages at the head of the inactive queue rather than the
2779 * tail, thus defeating the queue's LRU operation and ensuring that the
2780 * page will be reused quickly.
2782 _vm_page_deactivate(m, m->dirty == 0);
2786 * Grab a page, waiting until we are waken up due to the page
2787 * changing state. We keep on waiting, if the page continues
2788 * to be in the object. If the page doesn't exist, first allocate it
2789 * and then conditionally zero it.
2791 * This routine may sleep.
2793 * The object must be locked on entry. The lock will, however, be released
2794 * and reacquired if the routine sleeps.
2797 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2802 VM_OBJECT_ASSERT_WLOCKED(object);
2803 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
2804 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
2805 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
2807 if ((m = vm_page_lookup(object, pindex)) != NULL) {
2808 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
2809 vm_page_xbusied(m) : vm_page_busied(m);
2811 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
2814 * Reference the page before unlocking and
2815 * sleeping so that the page daemon is less
2816 * likely to reclaim it.
2818 vm_page_aflag_set(m, PGA_REFERENCED);
2820 VM_OBJECT_WUNLOCK(object);
2821 vm_page_busy_sleep(m, "pgrbwt");
2822 VM_OBJECT_WLOCK(object);
2825 if ((allocflags & VM_ALLOC_WIRED) != 0) {
2831 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2833 if ((allocflags & VM_ALLOC_SBUSY) != 0)
2838 m = vm_page_alloc(object, pindex, allocflags);
2840 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
2842 VM_OBJECT_WUNLOCK(object);
2844 VM_OBJECT_WLOCK(object);
2846 } else if (m->valid != 0)
2848 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
2854 * Mapping function for valid or dirty bits in a page.
2856 * Inputs are required to range within a page.
2859 vm_page_bits(int base, int size)
2865 base + size <= PAGE_SIZE,
2866 ("vm_page_bits: illegal base/size %d/%d", base, size)
2869 if (size == 0) /* handle degenerate case */
2872 first_bit = base >> DEV_BSHIFT;
2873 last_bit = (base + size - 1) >> DEV_BSHIFT;
2875 return (((vm_page_bits_t)2 << last_bit) -
2876 ((vm_page_bits_t)1 << first_bit));
2880 * vm_page_set_valid_range:
2882 * Sets portions of a page valid. The arguments are expected
2883 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2884 * of any partial chunks touched by the range. The invalid portion of
2885 * such chunks will be zeroed.
2887 * (base + size) must be less then or equal to PAGE_SIZE.
2890 vm_page_set_valid_range(vm_page_t m, int base, int size)
2894 VM_OBJECT_ASSERT_WLOCKED(m->object);
2895 if (size == 0) /* handle degenerate case */
2899 * If the base is not DEV_BSIZE aligned and the valid
2900 * bit is clear, we have to zero out a portion of the
2903 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2904 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
2905 pmap_zero_page_area(m, frag, base - frag);
2908 * If the ending offset is not DEV_BSIZE aligned and the
2909 * valid bit is clear, we have to zero out a portion of
2912 endoff = base + size;
2913 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2914 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
2915 pmap_zero_page_area(m, endoff,
2916 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
2919 * Assert that no previously invalid block that is now being validated
2922 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
2923 ("vm_page_set_valid_range: page %p is dirty", m));
2926 * Set valid bits inclusive of any overlap.
2928 m->valid |= vm_page_bits(base, size);
2932 * Clear the given bits from the specified page's dirty field.
2934 static __inline void
2935 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
2938 #if PAGE_SIZE < 16384
2943 * If the object is locked and the page is neither exclusive busy nor
2944 * write mapped, then the page's dirty field cannot possibly be
2945 * set by a concurrent pmap operation.
2947 VM_OBJECT_ASSERT_WLOCKED(m->object);
2948 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
2949 m->dirty &= ~pagebits;
2952 * The pmap layer can call vm_page_dirty() without
2953 * holding a distinguished lock. The combination of
2954 * the object's lock and an atomic operation suffice
2955 * to guarantee consistency of the page dirty field.
2957 * For PAGE_SIZE == 32768 case, compiler already
2958 * properly aligns the dirty field, so no forcible
2959 * alignment is needed. Only require existence of
2960 * atomic_clear_64 when page size is 32768.
2962 addr = (uintptr_t)&m->dirty;
2963 #if PAGE_SIZE == 32768
2964 atomic_clear_64((uint64_t *)addr, pagebits);
2965 #elif PAGE_SIZE == 16384
2966 atomic_clear_32((uint32_t *)addr, pagebits);
2967 #else /* PAGE_SIZE <= 8192 */
2969 * Use a trick to perform a 32-bit atomic on the
2970 * containing aligned word, to not depend on the existence
2971 * of atomic_clear_{8, 16}.
2973 shift = addr & (sizeof(uint32_t) - 1);
2974 #if BYTE_ORDER == BIG_ENDIAN
2975 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
2979 addr &= ~(sizeof(uint32_t) - 1);
2980 atomic_clear_32((uint32_t *)addr, pagebits << shift);
2981 #endif /* PAGE_SIZE */
2986 * vm_page_set_validclean:
2988 * Sets portions of a page valid and clean. The arguments are expected
2989 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2990 * of any partial chunks touched by the range. The invalid portion of
2991 * such chunks will be zero'd.
2993 * (base + size) must be less then or equal to PAGE_SIZE.
2996 vm_page_set_validclean(vm_page_t m, int base, int size)
2998 vm_page_bits_t oldvalid, pagebits;
3001 VM_OBJECT_ASSERT_WLOCKED(m->object);
3002 if (size == 0) /* handle degenerate case */
3006 * If the base is not DEV_BSIZE aligned and the valid
3007 * bit is clear, we have to zero out a portion of the
3010 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3011 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3012 pmap_zero_page_area(m, frag, base - frag);
3015 * If the ending offset is not DEV_BSIZE aligned and the
3016 * valid bit is clear, we have to zero out a portion of
3019 endoff = base + size;
3020 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3021 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3022 pmap_zero_page_area(m, endoff,
3023 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3026 * Set valid, clear dirty bits. If validating the entire
3027 * page we can safely clear the pmap modify bit. We also
3028 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3029 * takes a write fault on a MAP_NOSYNC memory area the flag will
3032 * We set valid bits inclusive of any overlap, but we can only
3033 * clear dirty bits for DEV_BSIZE chunks that are fully within
3036 oldvalid = m->valid;
3037 pagebits = vm_page_bits(base, size);
3038 m->valid |= pagebits;
3040 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3041 frag = DEV_BSIZE - frag;
3047 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3049 if (base == 0 && size == PAGE_SIZE) {
3051 * The page can only be modified within the pmap if it is
3052 * mapped, and it can only be mapped if it was previously
3055 if (oldvalid == VM_PAGE_BITS_ALL)
3057 * Perform the pmap_clear_modify() first. Otherwise,
3058 * a concurrent pmap operation, such as
3059 * pmap_protect(), could clear a modification in the
3060 * pmap and set the dirty field on the page before
3061 * pmap_clear_modify() had begun and after the dirty
3062 * field was cleared here.
3064 pmap_clear_modify(m);
3066 m->oflags &= ~VPO_NOSYNC;
3067 } else if (oldvalid != VM_PAGE_BITS_ALL)
3068 m->dirty &= ~pagebits;
3070 vm_page_clear_dirty_mask(m, pagebits);
3074 vm_page_clear_dirty(vm_page_t m, int base, int size)
3077 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3081 * vm_page_set_invalid:
3083 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3084 * valid and dirty bits for the effected areas are cleared.
3087 vm_page_set_invalid(vm_page_t m, int base, int size)
3089 vm_page_bits_t bits;
3093 VM_OBJECT_ASSERT_WLOCKED(object);
3094 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3095 size >= object->un_pager.vnp.vnp_size)
3096 bits = VM_PAGE_BITS_ALL;
3098 bits = vm_page_bits(base, size);
3099 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3102 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3103 !pmap_page_is_mapped(m),
3104 ("vm_page_set_invalid: page %p is mapped", m));
3110 * vm_page_zero_invalid()
3112 * The kernel assumes that the invalid portions of a page contain
3113 * garbage, but such pages can be mapped into memory by user code.
3114 * When this occurs, we must zero out the non-valid portions of the
3115 * page so user code sees what it expects.
3117 * Pages are most often semi-valid when the end of a file is mapped
3118 * into memory and the file's size is not page aligned.
3121 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3126 VM_OBJECT_ASSERT_WLOCKED(m->object);
3128 * Scan the valid bits looking for invalid sections that
3129 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3130 * valid bit may be set ) have already been zeroed by
3131 * vm_page_set_validclean().
3133 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3134 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3135 (m->valid & ((vm_page_bits_t)1 << i))) {
3137 pmap_zero_page_area(m,
3138 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3145 * setvalid is TRUE when we can safely set the zero'd areas
3146 * as being valid. We can do this if there are no cache consistancy
3147 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3150 m->valid = VM_PAGE_BITS_ALL;
3156 * Is (partial) page valid? Note that the case where size == 0
3157 * will return FALSE in the degenerate case where the page is
3158 * entirely invalid, and TRUE otherwise.
3161 vm_page_is_valid(vm_page_t m, int base, int size)
3163 vm_page_bits_t bits;
3165 VM_OBJECT_ASSERT_LOCKED(m->object);
3166 bits = vm_page_bits(base, size);
3167 return (m->valid != 0 && (m->valid & bits) == bits);
3171 * vm_page_ps_is_valid:
3173 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3176 vm_page_ps_is_valid(vm_page_t m)
3180 VM_OBJECT_ASSERT_LOCKED(m->object);
3181 npages = atop(pagesizes[m->psind]);
3184 * The physically contiguous pages that make up a superpage, i.e., a
3185 * page with a page size index ("psind") greater than zero, will
3186 * occupy adjacent entries in vm_page_array[].
3188 for (i = 0; i < npages; i++) {
3189 if (m[i].valid != VM_PAGE_BITS_ALL)
3196 * Set the page's dirty bits if the page is modified.
3199 vm_page_test_dirty(vm_page_t m)
3202 VM_OBJECT_ASSERT_WLOCKED(m->object);
3203 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3208 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3211 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3215 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3218 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3222 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3225 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3228 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3230 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3233 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3237 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3240 mtx_assert_(vm_page_lockptr(m), a, file, line);
3246 vm_page_object_lock_assert(vm_page_t m)
3250 * Certain of the page's fields may only be modified by the
3251 * holder of the containing object's lock or the exclusive busy.
3252 * holder. Unfortunately, the holder of the write busy is
3253 * not recorded, and thus cannot be checked here.
3255 if (m->object != NULL && !vm_page_xbusied(m))
3256 VM_OBJECT_ASSERT_WLOCKED(m->object);
3260 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3263 if ((bits & PGA_WRITEABLE) == 0)
3267 * The PGA_WRITEABLE flag can only be set if the page is
3268 * managed, is exclusively busied or the object is locked.
3269 * Currently, this flag is only set by pmap_enter().
3271 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3272 ("PGA_WRITEABLE on unmanaged page"));
3273 if (!vm_page_xbusied(m))
3274 VM_OBJECT_ASSERT_LOCKED(m->object);
3278 #include "opt_ddb.h"
3280 #include <sys/kernel.h>
3282 #include <ddb/ddb.h>
3284 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3286 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3287 db_printf("vm_cnt.v_cache_count: %d\n", vm_cnt.v_cache_count);
3288 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3289 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3290 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3291 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3292 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3293 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3294 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3297 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3301 db_printf("pq_free %d pq_cache %d\n",
3302 vm_cnt.v_free_count, vm_cnt.v_cache_count);
3303 for (dom = 0; dom < vm_ndomains; dom++) {
3305 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3307 vm_dom[dom].vmd_page_count,
3308 vm_dom[dom].vmd_free_count,
3309 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3310 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3311 vm_dom[dom].vmd_pass);
3315 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3321 db_printf("show pginfo addr\n");
3325 phys = strchr(modif, 'p') != NULL;
3327 m = PHYS_TO_VM_PAGE(addr);
3329 m = (vm_page_t)addr;
3331 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3332 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3333 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3334 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3335 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);